llms.txt is a proposed convention: a plain-text file at a site's root that summarises the site for large language models, in the way robots.txt gives crawlers instructions. It is widely recommended in AI-visibility content, including, until recently, on this site. The evidence does not really support the way it is usually described, and it is worth being straight about that.
What the major providers have actually said
Google's guidance on optimising for its AI features states that llms.txt is not needed for AI Overviews, AI Mode or its other generative search features. Google's own search advocates have said publicly that no Google Search system reads or acts on the file. That is about as direct as platform guidance gets.
More broadly, as of early 2026 no major model provider — Google, OpenAI, Anthropic, Meta — has publicly committed to reading llms.txt in their production search or answer systems. That is not the same as saying it is ignored everywhere, but it is a long way from the "new standard for AI visibility" framing the file usually receives.
Where it genuinely is used
The picture is not simply negative. Anthropic recommends the file in guidance aimed at writing documentation for agents. OpenAI uses it in the context of its Agents SDK and agentic commerce work. Chrome's Lighthouse added an agentic-browsing audit in 2026 that checks whether a site provides one.
The pattern is consistent: llms.txt has real traction in agent and developer tooling — software that reads a site in order to act on it — and very little demonstrated role in whether a brand gets cited in a consumer-facing AI answer. Those are different problems, and conflating them is where most of the confusion comes from.
What the adoption data shows
A large-scale crawl of around 300,000 domains in 2026 found roughly a 10% adoption rate, after about eighteen months of discussion. More usefully, analyses to date have not found a correlation between having the file and being cited more often by AI systems.
Absence of a measured correlation is not proof that it does nothing. But when a technique is recommended as essential, the burden of evidence sits with the recommendation, and that evidence has not arrived.
Why it gets over-recommended anyway
It is easy to explain, quick to implement, and produces a visible artefact a client can be shown. That combination makes it attractive to sell. The harder work — earning the independent coverage that actually gives an AI system a reason to trust a claim — is slower and cannot be demonstrated in an afternoon.
Be cautious of any AI-visibility proposal where the technical checklist is the substance of the offering.
So should you add one?
Reasonable position: add it if you want, expect little, and do not let it displace anything.
It costs almost nothing to publish. If your product has developer or agent-facing documentation, there is a defensible reason to have one. What it should not do is occupy a line item in a strategy where it substitutes for the work that moves citation: consistent, verifiable facts about your company, clean entity information, structured data, and independent sources saying something about you.
This site publishes one, for the agent-tooling reason rather than because it expects Google to read it.
What to do instead, in order
- Make the factual claims about your company consistent everywhere they appear.
- Give AI systems clean, unambiguous entity information and structured data.
- Answer the questions buyers actually ask, in plain language, on pages that are easy to parse.
- Earn independent coverage, because corroboration is the signal that is genuinely scarce.
Guidance in this area changes quickly. If Google or another provider commits to reading llms.txt in production, this advice changes with it — but write to the evidence available, not to the ambition.
See our AI visibility service for how the structural work and the earned-media work run together, or what generative engine optimisation actually involves.